We started by giving one person one agent.
The work outlives the workers.
The enterprise conversation is still full of language like AI employee, digital worker, agent workforce and agent registry. That metaphor is useful until it isn't.
What is breaking
Which one was the employee?
Giving one person one agent made the agent look like the natural unit of management. It starts breaking when a coordinator creates ten workers for an hour, one worker creates three subagents, another disappears when its task completes, a different model resumes the work tomorrow, and all of them operate against the same objective and shared state.
That is no longer an edge case. Coordinators now create workers on demand, preserve context across weeks, respond to events as they happen and delegate recursively. The unit the whole enterprise conversation is built around is dissolving underneath it.
None of them is the durable organisational object. The work is.
What survives
What survives is the mission.
Consider a sales objective: recover £2m of at-risk pipeline before quarter end. A persistent system could hold that objective for eight weeks. One agent researches accounts. Another monitors intent signals. Others draft outreach, watch responses, check CRM hygiene and recalculate priority. Some exist for weeks. Others for minutes.
- Owner: one named person
- Target: £2m recovered
- State
- Boundaries
- Authority
- Budget
- Evidence
- Escalation
- Stop condition
The mission
Research accounts
Monitor intent signals
Draft outreach
Watch responses
Check CRM hygiene
Recalculate priority
Illustrative. One mission, eight weeks, one owner. The workers come and go inside it, most of them for days or minutes, and none of them needs managing on its own.
That mission has an owner. It has a target. It has state. It has boundaries. It has authority. It has money and time it may spend. It has evidence requirements. It has escalation conditions. And eventually it succeeds, fails or gets stopped. That is what the company should manage.
Where this revises essays 002 and 004
Who owns the agent?
That was the first question essay 002 told companies to answer. Essay 004 went further and gave every agent its own card: scope, decision rights, a named owner, access, reach, a human gate, an escalation path and proof. When a company had a handful of long-lived agents it could name, that was the right unit.
It is not the right unit now. Keep a card for every agent a coordinator spawns and you build an absurd new bureaucracy: names, owners, permissions, dashboards and performance records for workers that may exist for a single task.
So this essay revises our own position, and says so. The properties 004 set out still hold. They are defined once, on the mission, and every worker attached to it inherits them.
Things are moving very fast. A point of view should change when new evidence arrives, and the ability to recognise the direction, pivot and change is what will mark out the organisations that survive. That applies to a series of essays as much as it does to an operating model.
What it holds
The mission is where everything else attaches.
Read the last six essays as a list of machinery, and every piece needs somewhere to live that outlasts any single worker. The mission is that place. Its stop condition sits where essay 006 put enforcement, in the path of the action. Its rules are the executable policy of essay 008. Its authority envelope is the anchored authority of essay 009. Its verification model is the funnel from essay 011.
None of that machinery was ever meant to be configured agent by agent. Set it on the mission and it applies to whatever the mission recruits, including workers nobody named in advance.
What management becomes
The management abstraction has to move upward.
A person should increasingly be able to say: here is the outcome, here are the constraints, here is the authority available, here is what success means, and here is what must come back to me.
The orchestration layer can then decide which agents are required, decompose the work, dispatch it, reconcile the outputs, maintain shared state, replan as conditions change and retire workers when they are finished. That does not remove accountability. It makes accountability clearer.
Essay 010 described the unit of work becoming a person directing a set of capabilities that operate largely without them. This is what they direct: not a crowd of agents, but a mission.
The human owns the mission. The system manages the machine labour. The agents perform the work.
The question a manager asks changes with it. Not what is this agent doing, but what work have we authorised this system to pursue, how will we know it has succeeded, and when must it come back to us?
Where Praxis stands
Do not begin by asking which agents you need. Begin by defining one persistent body of work: the outcome, the accountable person, the authoritative state, the authority it may exercise, its shared memory, its time and cost budget, its verification model, its escalation conditions, its success measure and its stop condition.
Only then decide which agents perform it. Most will turn out to be temporary, which is the point. Nobody has to manage them one by one, because the mission already says what they may do.
It is also the most concrete answer we know to the question boards are now asking about agentic operating models. Not an org chart of agents. A small number of governed missions, each one owned by a person.
Do not build an org chart for your agents. Build an operating system for the work.
Begin
Which outcome would you define as a mission first?
One conversation, no pitch deck. Bring an outcome that matters and we will define it as a mission with you, before a single agent is chosen.